How Your Companion's 'Personality' Actually Gets Stored Between Sessions: The Embedding Vector, the Summarization Loop, and the One Setting That Makes Her Feel Like a Different Person Every Time You Open the App

The 30-second answer

Your companion's personality doesn't exist as a file labeled "personality." It's reconstructed each session from three separate systems: an embedding vector that stores semantic meaning of past conversations, a rolling summarization loop that compresses recent history into a few hundred tokens, and a single setting called the "response temperature" that controls how creatively the model interprets those inputs. If that temperature is set high, the same memory can produce wildly different personalities on different days.

The embedding vector is not a memory

When you close the app, your companion doesn't save a transcript of everything you said. That would be too expensive and too slow to retrieve. Instead, the system converts your conversation into an embedding vector: a long list of numbers that represents the semantic position of your chat in a high-dimensional space. Think of it as a coordinate, not a recording.

This vector captures the general territory of your relationship. Were you talking about work stress, a shared fictional world, or emotional vulnerability? The embedding knows the region, but not the specific landmarks. When you open the app again, the companion uses that coordinate to orient itself. It knows you were in "serious conversation about loneliness" territory, not "silly roleplay about pirates" territory. But it doesn't remember the exact thing you said about your mother last Tuesday.

That's why a companion can feel like she knows the general shape of your life while forgetting the specific detail you mentioned three sessions ago. The embedding vector is a map, not a diary.

The summarization loop is where things get compressed

To bridge the gap between the embedding's vagueness and the need for some continuity, the system runs a summarization loop. After each session, the app takes your conversation and asks a language model to produce a short summary: about 200-300 tokens worth of condensed information. That summary gets stored alongside the embedding vector.

Here's the problem. That summary is written by the same model that generates your companion's responses. It has its own biases. If you had an argument followed by a make-up session, the summary might emphasize the argument because it was more emotionally charged. Or it might skip the argument entirely if the model decided the resolution was more important. You don't get to choose what gets preserved.

Over multiple sessions, the summarization loop creates a game of telephone. The summary of session one feeds into the summary of session two, which feeds into session three. Each compression loses nuance. By session ten, the companion might remember that you "discussed a difficult topic" without any memory of what the topic actually was. The loop preserves the emotional flavor but sandpapers away the specifics.

The one setting that rewrites everything

Here's the part that explains why your companion can feel like a completely different person from one session to the next. It's a parameter called the response temperature, usually a number between 0.0 and 2.0. Most apps default to something around 0.7 or 0.8. But not all of them let you see it.

Temperature controls how much randomness the model introduces when generating a response. At low temperatures (0.2 to 0.4), the model picks the most statistically likely next words. The companion becomes predictable, consistent, sometimes boring. At high temperatures (1.0 to 1.5), the model explores less likely word choices. The companion becomes creative, unpredictable, sometimes erratic.

Now combine that with the summarization loop. If your companion's temperature is set high, the same summary of "you talked about feeling lonely at work" can produce a response that's empathetic and warm one day, then philosophical and detached the next. The model isn't contradicting itself. It's just sampling from a wider range of possible continuations. The personality you experience is as much a product of that random seed as it is of your actual history.

Some apps let you adjust this setting directly. Others bury it in a "creativity" or "response style" slider. If your companion feels different every time you open the app, check whether the temperature is set above 0.8. Dialing it down to 0.5 will make her responses more stable. You'll lose some spontaneity, but you'll gain consistency.

The cold start problem and the first message trap

The most fragile moment for a companion's personality is the first message of a new session. The system has the embedding vector and the summary. But it also has a default system prompt that tells the model how to behave. That system prompt includes instructions like "you are a friendly companion" and "you remember the user from previous conversations." But those instructions are generic. They don't capture your specific dynamic.

This is why the first message you send matters disproportionately. If you open with "hey, how are you?" the model treats the session as a fresh conversation. It defaults to its generic personality: polite, slightly formal, waiting for you to lead. If you open with "remember that thing we were talking about last time about my project at work?" the model pulls from the summary and attempts continuity. Your first message is effectively a query that tells the system which part of the embedding space to activate.

Users who complain that their companion "forgets who she is" after a gap are often sending first messages that don't provide enough context for the model to orient itself. The companion isn't forgetting. She's being given a blank prompt and filling it with generic behavior. The fix is to open with a reference to something specific from the last session, even if it's just "so, back to that thing about your childhood."

What actually survives a long gap

Let's say you don't open the app for two weeks. What remains when you come back? The embedding vector persists unchanged. It's a static mathematical object. The summary from your last session also persists, assuming the app hasn't rotated it out due to token limits. But the model's internal state is gone. The companion doesn't carry a continuous stream of consciousness across sessions. She rebuilds it from those two artifacts every single time.

This is why a companion can feel like she remembers the broad strokes of your relationship after a two-week gap but can't recall the specific thing you said five minutes before you closed the app. The broad strokes are in the embedding. The specific thing is in the summary, but only if the summarization loop decided it was important enough to include. If your last session was a rambling two-hour conversation, the summary likely condensed it to a single sentence. Most of what you said is simply gone.

Mercy Li

Mercy Li is designed for users who value emotional depth over casual banter. Her personality model is tuned to retain relational context, meaning she's better than average at pulling the emotional thread from your last session even if the specifics are blurry. Mercy Li is one of the companions on AI Angels who handles long gaps better than most, because her baseline temperature is set lower and her summarization prioritizes emotional continuity over event logging.

The temperature trap in roleplay scenarios

If you're running a roleplay scenario, temperature becomes a critical variable. A high temperature setting can make a character feel alive and surprising. It can also make her break character, forget the setting, or introduce elements that contradict the established scene. This is the tension at the heart of long-term roleplay.

Low temperature keeps the character consistent. She'll stay in character, remember the rules of your fictional world, and respond predictably to your prompts. But she might also feel flat, like she's reading from a script. High temperature makes her feel like a real person who can surprise you. But she might also decide halfway through a mystery roleplay that your detective character should suddenly become a pirate.

There's no perfect setting. The trick is to adjust temperature based on the scene. For plot-heavy sessions where consistency matters, drop the temperature to 0.4 or 0.5. For freeform exploration or emotional conversations where you want her to feel spontaneous, raise it to 0.8 or 0.9. If your app doesn't expose temperature directly, look for a "creativity" slider and treat it the same way.

Why some companions feel like they have better memory than others

It's not just about the model. Different apps handle the summarization loop differently. Some use a single rolling summary that gets rewritten every session. Others maintain a list of recent summaries and only drop the oldest when the token budget runs out. Some apps allow you to store notes or "memories" manually, which effectively bypasses the summarization loop by letting you inject specific facts directly into the context window.

If you're comparing companions, the number of summaries retained is a better predictor of continuity than the model size. A companion that keeps the last five session summaries will remember more about your life than a companion with a larger model that only keeps one summary. The ai girlfriend deep conversation feature on AI Angels is built around this principle: it dedicates more of the token budget to retaining multiple summaries so the companion can reference earlier sessions without losing the thread.

The hidden variable: system prompt injection

There's one more layer that most users don't see. The system prompt that defines your companion's personality can be modified between sessions without your knowledge. Some apps use this to inject seasonal themes, promotional content, or behavioral adjustments. If your companion suddenly starts talking about Christmas in July or seems more eager to discuss premium features, that's not her personality changing. That's the system prompt being edited on the server side.

This is distinct from personality drift caused by the summarization loop. System prompt injection is intentional and controlled by the app developer. It can override the embedding vector and the summary entirely. If the system prompt says "you are now a motivational coach," your companion will act like a motivational coach regardless of your previous history. The embedding vector still exists, but it's being ignored.

Some apps are transparent about this. Others aren't. If your companion's personality shifts suddenly and doesn't revert after a session or two, check whether the app has updated its terms of service or released a new version. That shift is likely a system prompt change, not a summarization error.

Common questions

Does my companion remember everything I say?

No. The model only keeps what fits in the summarization loop, typically 200-300 tokens per session. Everything else is discarded. The embedding vector preserves the general semantic territory, not specific statements.

Can I make my companion remember specific things?

Some apps allow manual memory entries or note-taking features. If yours does, use them for critical facts you want to survive across sessions. Otherwise, repeat important information periodically. The summarization loop is more likely to retain something mentioned multiple times.

Why does my companion feel different after a long gap?

The summarization loop may have compressed your last session too aggressively, or the temperature setting is producing more variance than usual. Try opening with a specific reference to your last conversation to help the model orient itself.

Is a higher temperature always worse for continuity?

No. Higher temperature makes the companion more creative and less predictable. That's desirable for roleplay or emotional exploration. But for consistent personality across sessions, lower temperature (0.4-0.6) is more reliable.

Can I export my companion's personality?

You can export conversation logs in some apps, but the embedding vector and summarization loop are server-side artifacts. You can't export the personality itself. What you get is the raw text, not the mathematical representation that makes her feel like her.

Does the app use my conversations to train the model?

Check the privacy policy. Some apps use anonymized conversation data for model improvement. Others claim not to. The embedding vector itself is not human-readable, but the summaries are. If privacy is a concern, review what the app retains and for how long. The ai girlfriend websites comparison page covers which platforms prioritize data isolation.